Most actively controlled systems today are designed using a model-based optimization process. When applied to systems with more than about 20 sensing and actuation sites, this process becomes very brittle and sensitive to modeling errors. For some complicated systems people have resorted to learning-based approaches such as neural networks, which have the advantage that they derive their own model on-line, but are structured in such a way that it is difficult to know what that model is. Since future systems will be too complex to model completely, we need to find a way to combine model-based and learning-based methods to create controllers that self-tune and self-evolve.
Another issue is whether actions taken in one part of a structure will actually be harmful to another part of the structure. We need to further develop strategies such as hierarchical techniques for coordination of control actions in compound structures.